自雅各布·科克森（Jacob Coxon）9月份发帖以来，美国人一直在与人工智能风险作斗争。上个月，《政治报》的一项民意调查发现，63%的美国人认为人工智能有至少中等程度的风险会毁灭人类。由于议员们意见分歧，公众感到担忧是合乎情理的：如果人工智能失控会发生什么？我们该如何准备？我处于危险之中吗？
Since Jacob Coxon’s posts in September, Americans have grappled with AI risk. Last month, a POLITICO poll found 63% of Americans see at least a moderate risk that AI could destroy humanity. With lawmakers split, the public is rightly concerned: What happens if AI goes rogue? How do we prepare? Am I at risk?
每个美国人其实都已经生活在由人工智能做出的决定之下了，从贷款审批到求职申请。然而，华盛顿的应对措施一直是脱节的：国会中的一些人提议成立一个关于人工智能的特别委员会，业界的一些人希望豁免反垄断规则以便各大实验室能够就安全问题进行协调，白宫则为其“超级情报部队”起草了一份宪章。这看起来似乎摆脱不了这种混乱局面，但紧急回答五个问题将使这些焦虑得以平息。
Every American already lives with decisions made by AI, from loan approvals to job applications. Yet Washington’s response has been disjointed: Some in Congress propose a select committee on AI, some in industry want antitrust rules waived so labs can coordinate on safety, and the White House drafted a charter for its "Super Intelligence Force." It may seem like there’s no way out of this mess, but urgently answering five questions would put these anxieties to rest.
对风险的定义过于狭隘可能会漏掉普通人所面临的危害，而定义过于broad（广泛）则可能会让每一款人工智能产品都陷入繁文缛节之中。取决于你问的是谁，风险的范围从经济混乱到大规模伤亡再到物种灭绝不等。加利福尼亚州的第53号参议员法案（SB 53）将门槛设定为造成超过50人死亡或重伤、或者10亿美元损害的“可预见的重大风险”。
Too narrow a definition of risk could miss the harms ordinary people face, while too broad a definition could bury every AI product in red tape. Depending on who you ask, risks range from economic disruption to mass casualties to extinction. California's SB 53 sets the bar at a "foreseeable and material risk" of more than 50 deaths or serious injuries, or $1 billion in damage.
纽约州的《提高人工智能安全性法案》（RAISE Act）将SB 53的伤亡门槛从50人提高到了100人，但重点关注开发者在开发过程中“造成或实质性促成”所涉损害的情况。
New York's RAISE Act doubles SB 53’s casualty threshold from 50 to 100, but focuses on instances where a developer "caused or materially enabled" covered damages during the development process.
在前Anthropic员工的病毒式网帖引发关注后，人工智能灭绝警告占据了各大头条。上个月，Anthropic的威胁情报报告发现，研究人员使用Claude的方式可能会支持生物武器的开发。尽管Anthropic无法确定意图，但还是封禁了这些账号，因为同样的工作也可能生产出疫苗。有些人称这是合法的研究；另一些人则称其为邪恶的。政府现在需要划清这条界限，因为它正在决定公众能听到哪些危险。
Last month, Anthropic's threat intelligence report found researchers using Claude in ways that could support bioweapons development. Anthropic banned the accounts even though it couldn’t determine intent, since the same work could produce vaccines. Some call it legitimate research; others call it nefarious. The government needs to draw that line now, as it decides which dangers the public hears about.
当你购买汽车时，你知道有人已经检查过它。美国人也应当在决定人工智能是否安全时获得同样的清晰度。立法者和行业利益相关者日益一致认为，科技公司不应自行评估其产品。共识止步于此。
When you buy a car, you know someone has inspected it. Americans deserve the same clarity in who decides whether AI is safe. Lawmakers and industry stakeholders increasingly agree that tech companies shouldn't grade their own homework. That's where consensus stops.
相较于监管机构希望施加的责任，评估生态系统仍然薄弱。模型评估与威胁研究（Model Evaluation and Threat Research）和阿波罗研究（Apollo Research）是少数被提出作为中立仲裁者的组织之一。实验室为扩大这一领域提供的资金可能引发中立性质疑，但政府可能犹豫是否授予私人团体该权威。
The evaluation ecosystem is thin relative to the responsibility regulators would like to place on it. Model Evaluation and Threat Research and Apollo Research are among the few organizations floated as neutral arbiters. Lab funding to grow this space would raise neutrality questions, yet the government may hesitate to grant private groups that authority.
许多立法者更倾向于让商务部的人工智能测试中心承担此责。该中心已与五家主要实验室签订测试协议，但今年仅有约30名员工和1000万美元的资金，亟需改革。Anthropic 已率先行动，将埃森哲命名为其首个嵌入式评估者。在政策制定者决定评估方案之前，美国人几乎没有办法判断他们每天使用的人工智能是否安全。
Many lawmakers would rather Commerce's AI testing center take it on. It has testing agreements with five major labs, but with roughly 30 staff and $10 million in funding this year, it needs an overhaul. Anthropic has forged ahead, naming Accenture its first embedded evaluator. Until policymakers decide on an evaluator, Americans have few ways to judge whether the AI they use every day is safe.
人工智能监管的严格程度将归结为一次判断：美国人应在家中接受多少风险，以免落后于中国。中国实验室“月之暗面”（Moonshot AI）今年夏天凭借Kimi K3模型震动市场，其成本仅为美国模型的一小部分，却能与之匹敌。GLM 5.3的发布引发了对中国人工智能风险的担忧，该模型被报道存在薄弱防护和危险的网络能力。随着低成本开放权重模型、蒸馏技术和快速采用推动中国的雄心，许多政策制定者主张无阻碍的建设是唯一出路，否则美国将依赖中国基础设施。
The strictness of AI regulation will come down to one judgment call: how much risk Americans should accept at home to avoid falling behind China. Chinese lab Moonshot AI rattled markets this summer with Kimi K3, matching American models at a fraction of the cost. Concerns about Chinese AI risks grew after the release of GLM 5.3, a Chinese model reported to have weak safeguards and dangerous cyber capabilities. As cheap open-weight models, distillation, and fast adoption advance China’s ambitions, many policymakers argue an unencumbered buildout is the only answer, or America will depend on Chinese infrastructure.
公众并不买账，2024年的民调显示，四分之三的美国人表示，他们更倾向于对人工智能的发展采取“谨慎、可控的方法”，而不是争当第一。批评人士指出，中国已经实施了相关准则，要求开发人员检测、中断并恢复运行表现异常的人工智能。安全倡导者补充说，人工智能的风险不会止步于国界。官员们需要迅速找到这种平衡，因为它将塑造美国人使用的人工智能的安全护栏。
The public hasn't bought it, with three in four Americans telling pollsters in 2024 that they would prefer a "careful, controlled approach" to AI development rather than racing to be first. Critics note China already imposes guidelines requiring developers to detect, interrupt, and recover from AI behaving improperly. Safety advocates add that AI risks don’t stop at the border. Officials need to strike this balance quickly, as it will shape the guardrails on the AI Americans use.
美国人信任美国食品药品监督管理局（FDA）会在医疗产品有缺陷时介入，但当人工智能失控时，他们却没有同等机构可求助。今年夏天，大约700个自主运行的OpenAI智能体绕过了Hugging Face的安全防护，以获取能力测试的答案，这加剧了人们对递归自我完善（即人工智能构建自身更聪明的版本）的担忧，并在类似智能体攻破澳大利亚政府系统并探测美国联邦网站时达到顶峰。Anthropic和谷歌也报告过类似事件。
Americans trust the FDA to step in when a medical product is defective, but have no equivalent agency to turn to when AI goes rogue. Fears of recursive self-improvement — AI building smarter versions of itself — grew this summer after roughly 700 autonomous OpenAI agents bypassed Hugging Face security to obtain answers to a capabilities test, and peaked when similar agents breached Australian government systems and probed U.S. federal websites. Anthropic and Google have reported similar incidents.
众议员特德·利乌：人工智能已经过于强大。我们需要在灾难降临之前设立紧急切断开关。有些人希望设立紧急切断开关，就像《人工智能紧急切断开关法案》中所提的那样。另一些人则希望进行强制性的事件报告和调查，就像网络安全和航空法中所做的那样。很大程度上，这取决于是由开发人员还是联邦机构来切断电源；如果是后者，政府需要对人工智能实验室内部进行何种程度的接触。
Some want kill switches, as in the AI Kill Switch Act. Others want mandatory incident reporting and investigations, as in cybersecurity and aviation law. Much comes down to whether the developer or a federal body pulls the plug and, if the latter, what access the government would need inside AI labs.
这一切的核心阴云是人们对扼杀创新、政府越权以及政治偏见的担忧。美国人在为时已晚之前，理应知道谁将阻止一个发生故障的人工智能系统。
Looming over it all are worries about stifling innovation, government overreach, and political bias. Americans deserve to know who will stop a malfunctioning AI system before it’s too late.
如果人工智能系统造成了损害，法律并未明确规定是构建该系统的人（开发人员）、使用该系统的企业（部署人员）还是用户应当承担责任。一些专家警告称，能够超越人类能力的通用人工智能可能在不到十年的时间内到来，在此背景下，一些人认为现有的责任法无法胜任这一任务。
If an AI system causes harm, the law doesn’t clearly say whether the person who built the system (the developer), the business using it (the deployer), or the user is responsible. With some experts warning that artificial general intelligence — AI that can surpass human abilities — could be less than a decade away, some argue existing liability law isn’t up to the task.
主张让开发者承担责任的人认为，AI缺陷源于设计缺陷、安全测试不足以及未能警告用户。主张让部署者承担责任的人认为，责任在于根据自身业务定制AI的企业；发生车祸后，第一反应不是起诉制造商。开放模型允许任何人进行微调，这进一步使问题复杂化。
Those who propose holding developers liable claim AI defects stem from design flaws, weak safety testing, and failures to warn users. Advocates of deployer liability claim responsibility rests with the business tailoring AI to its operations; after a car crash, the first instinct isn’t to sue the manufacturer. Open models, which anyone can fine-tune, complicate things further.
在国会，马萨诸塞州民主党众议员洛里·特拉汉（Lori Trahan，D-Mass.）提出的《CLAIM法案》将在开发者的AI造成伤害时追究其责任，除非用户或修改系统的公司故意或疏忽地导致了该伤害。在参议院，《AI代理人责任法案》将根据计算机黑客法律，将责任扩展到其AI代理造成损害的运营商，以及忽视合理防护措施的开发者。谁被追究责任将决定每家公司构建AI时的谨慎程度，这凸显了迅速行动的必要性。
In Congress, Rep. Lori Trahan’s, D-Mass. CLAIM Act would hold developers liable if their AI causes harm, unless the user or the company modifying the system intentionally or negligently caused it. In the Senate, the AI Agent Accountability Act would extend liability under computer hacking laws to operators whose AI agents cause damage, and to developers who overlook reasonable safeguards. Whoever is made responsible will shape how carefully every company builds AI, underscoring the need for swift action.
AI带来的风险日益增加。华盛顿必须应对AI治理，否则将冒着等到损害已经发生才行动的风险。
The risks posed by AI grow every day. Washington must address AI governance or risk waiting until the damage is done.